A multi-organ segmentation method for abdominal CT images based on deep learning

Abstract

The present invention discloses a method for multi-organ segmentation of abdominal CT images based on deep learning. The method is specifically implemented as follows: (1) constructing a training dataset containing abdominal CT images and their corresponding multi-organ segmentation results; (2) designing a segmentation network based on a dual self-attention mechanism and multi-scale feature fusion; (3) constructing a network loss function by combining Dice loss and Focal loss; (4) training the network using the training dataset; and (5) using the trained network to segment the various organ regions in the abdominal CT images. By adopting the dual attention and multi-scale feature fusion mechanisms, the present invention can establish long-range dependencies in a more targeted manner while introducing a very small number of parameters and floating-point calculations, thereby solving the problem of low segmentation accuracy of irregularly shaped long organs.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a multi-organ segmentation method for abdominal CT images based on deep learning. Background Art

[0002] With the continuous advancement of medical imaging equipment and intelligent medical technology, medical image segmentation has become an indispensable step in computer-assisted diagnosis (CAD). Multi-organ segmentation in abdominal computed tomography (CT) imaging facilitates tasks such as organ or tissue morphological analysis, disease diagnosis, lesion localization, and treatment plan formulation, improving the accuracy and efficiency of CAD. Currently, clinical multi-organ segmentation in the abdomen relies primarily on manual delineation by experts. However, a typical CT scan contains hundreds of two-dimensional slices. Manual segmentation of each slice is not only time-consuming and labor-intensive, but also can reduce segmentation accuracy due to prolonged manual work. The size, morphology, and position of various tissues and organs, especially luminal organoids such as the pancreas and esophagus, vary significantly across slices. The large individual differences between organs, as well as the proximity and blurred boundaries between organs, pose significant challenges to multi-organ segmentation in abdominal CT sequences. Therefore, developing accurate and rapid automatic segmentation methods for multi-organs in the abdomen has become a research hotspot in recent years.

[0003] With the recent advancements in artificial intelligence and machine vision, medical image segmentation methods based on neural networks and deep learning have become mainstream. Convolutional Neural Networks (CNNs), with their powerful nonlinear extraction capabilities, have achieved remarkable results in computer vision. Most CNN-based segmentation methods improve performance by simply deepening the network or stacking a large number of local convolutions. However, this approach fails to effectively expand the network's receptive field and is not well suited for complex abdominal CT images with numerous organs. Summary of the Invention

[0004] In response to the shortcomings and deficiencies of the existing technology, the present invention integrates dual self-attention and multi-scale feature fusion mechanisms into the construction of a U-shaped deep convolutional neural network, aiming to provide a multi-organ segmentation method for abdominal CT images based on deep learning. While introducing a very small number of parameters and floating-point calculations, it establishes long-distance dependencies in a more targeted manner, solving the problem of low segmentation accuracy of irregularly shaped long organs.

[0005] A multi-organ segmentation method for abdominal CT images based on deep learning, comprising the following steps:

[0006] (1) Establish a training dataset A containing abdominal CT images and their corresponding multi-organ segmentation results;

[0007] (2) Construct a segmentation network based on dual self-attention mechanism and multi-scale feature fusion, called DAMF-Net, which specifically includes:

[0008] (2-a) A U-shaped structure with skip connections is used as the network benchmark framework, in which the encoding path and decoding path are composed of 5 cascaded encoding blocks and decoding blocks respectively. In order to avoid redundant information transmission and enhance the recognition ability of small targets and edges, a skip connection with a residual block is added between the encoding and decoding blocks of the same level. In addition, in order to compensate for the information loss caused by the limited receptive field, establish the connection between global information, and lightweight the network, a feature extraction module consisting of a multi-scale information fusion module and a dual self-attention module in parallel is introduced at the deepest layer of the network. The multi-scale information fusion module is abbreviated as MSCF and the dual self-attention module is abbreviated as DSA.

[0009] (2-b) The coding block described in step (2-a) is characterized in that it includes a coding main path and a residual branch, and in the coding main path, the input feature map is first subjected to two 3×3 convolutions in sequence, the first 3×3 convolution layer is sequentially connected to a normalization layer and a ReLU activation layer, and the second 3×3 convolution layer is sequentially connected to a normalization layer; in order to prevent network degradation, a residual branch is added to the coding block to perform 1×1 convolution and normalization on the input features; then, the results of the residual branch and the coding main path are added, and finally, the output of the coding block is obtained through a ReLU activation layer and a 2×2 maximum pooling layer;

[0010] (2-c) The decoding block described in step (2-a) is characterized in that it includes a decoding main path and a residual branch, and in the decoding main path, the input feature map is first subjected to two 3×3 convolutions in sequence, the first 3×3 convolution layer is connected in sequence to a normalization layer and a ReLU activation layer, and the second 3×3 convolution layer is connected to a normalization layer; in order to prevent network degradation, a residual branch is added to the encoding block, and the input features are subjected to 1×1 convolution and normalization; then, the residual branch is added to the result of the decoding main path, and finally the output of the encoding block is obtained through a ReLU activation layer and upsampling, wherein the upsampling is implemented by 2×2 deconvolution;

[0011] (2-d) The residual block described in step (2-a) is characterized in that the specific structure includes: taking the output of the current level and the next level coding block as input, wherein the output of the current level coding block is denoted as F i , the output of the next level coding block is recorded as F i+1 ; The residual block first converts F i+1 Perform 2×2 deconvolution and ReLU activation operations to obtain the same value as F i Feature maps F of the same size up , and Fi With F up Subtract and obtain the residual feature map F containing small targets and edge details sub =F i -F up ; In order to further enhance the expression of residual features, F sub Perform two consecutive 3×3 convolution operations, and connect a normalization layer and a ReLU activation layer in sequence after each 3×3 convolution layer;

[0012] (2-e) The multi-scale information fusion module described in step (2-a), namely, MSCF, is characterized by the following specific structure: first, the input feature F in Perform 1×1 convolution, normalization, and ReLU activation operations in sequence to obtain feature F conv1×1 The purpose of this operation is to reduce the number of channels, reduce the number of parameters, and speed up the calculation process in the subsequent modules; then, F conv1×1 Input them into the pyramid pooling module to obtain the features F containing local information MS1 , and F conv1×1 Input to the anisotropic strip pooling module to obtain the feature F containing long-distance context information MS2 ; F MS1 and F MS2 Perform addition fusion, and then perform 1×1 convolution and normalization operations in sequence to obtain F that fuses multi-scale feature information MSCF ; In order to ensure the semantic consistency, the input feature F in With F MSCF The sum is added and output through the ReLU activation layer, and then 3×3 convolution, normalization, and ReLU activation operations are performed in sequence. Finally, the channel dimension is restored to the size of the input feature through 1×1 convolution to obtain the output feature F out ;

[0013] (2-f) The pyramid pooling module described in step (2-e) is denoted as PPM, which is characterized by the following structure: first, the input feature F conv1×1 Perform average pooling with kernels of 3×3, 5×5, and 7×7 respectively to obtain feature maps F of different scales pool3×3 、F pool5×5 and F pool7×7 , and respectively for F pool3×3 、F pool5×5 and F pool7×7 Perform 3×3 convolution and normalization in sequence to obtain the feature map F p1 、F p2 、F p3 , then, respectively for F p1 、F p2 、F p3Perform deconvolution with kernels of 3×3, 5×5, and 7×7 to obtain features F1, F2, and F3; finally, add and fuse F1, F2, and F3, and perform ReLU activation, 3×3 convolution, normalization, and ReLU activation in sequence to obtain feature F MS1 :

[0014] (2-g) The anisotropic strip pooling module described in step (2-e) is denoted as ASPM, which is characterized by the following structure: first, using pooling kernels of size H×1 and 1×W to pool F in the vertical and horizontal directions respectively. conv1×1 Perform average pooling to generate feature map F h and F v , where H and W represent the input feature map F conv1×1 The height and width of F h Perform 1×3 convolution and normalization in sequence to obtain the feature F′ h , for F v Perform 3×1 convolution and normalization in sequence to obtain the feature F′ v , and the obtained feature F′ h and F′ v Copy and expand along the vertical and horizontal directions respectively to obtain the same feature as the input F conv1×1 The feature maps F1′ and F2′ of the same size are then added and fused to obtain the feature F containing the long-range context information relationship. fusion ; Finally, for F fusion Perform ReLU activation, 3×3 convolution, normalization, and ReLU activation in sequence to obtain feature F MS2 ;

[0015] (2-h) The dual self-attention module described in step (2-a) is denoted as DSA, which is characterized by a specific structure comprising: a position attention module and a channel attention module in parallel, which takes the input feature F down Input the position attention module and channel attention module respectively to obtain the feature F PAM and F CAM , and then add and fuse them to get the output feature

[0016] (2-i) The position attention module described in step (2-h), denoted as PAM, is characterized by a specific structure including:

[0017] (2-iI) For input features Where W′ and H′ are the width and height of the feature map, C is the number of channels of the feature map, and F is calculated on the three branches. down Perform shift convolution to obtain features and

[0018] (2-i-Ⅱ) In order to obtain the mutual dependence between pixels, the reshape operation is used to transform F p1 and F p2 Reconstructed into and And F′ p1 and F′ p2 Perform matrix multiplication and Softmax activation to obtain the weight matrix

[0019] (2-i-Ⅲ) will Refactored to And F′ p3 With M PAM Perform matrix multiplication to obtain

[0020] (2-i-Ⅳ)will Refactored to And F′ pmap With input feature F down Perform additive fusion and obtain output features through shift convolution

[0021] (2-j) The channel attention module described in step (2-h), denoted as CAM, is characterized by its overall structure being similar to that of PAM, specifically comprising:

[0022] (2-j-Ⅰ) Use reshape operation to reshape F on three branches respectively. down Reconstructed into

[0023]

[0024] (2-j-Ⅱ) for F c1 Transpose and obtain And F′ c1 and F c2 Perform matrix multiplication and Softmax activation to obtain the weight matrix

[0025] (2-j-Ⅲ) F c3 With M CAM Perform matrix multiplication to obtain

[0026] (2-j-Ⅳ) Refactored to And F′ cmap With input feature F down Perform additive fusion and obtain output features through shift convolution

[0027] (3) Combining Dice loss and Focal loss to construct the loss function L of the DAMF-Net network:

[0028] L=L focal +λL dice

[0029]

[0030]

[0031]

[0032] Among them, L focal and L dice Represent Focal loss and Dice loss respectively, λ is the weight coefficient, which controls the relative importance of Focal loss and Dice loss. λ is preferably a constant between 0.3 and 0.7. represents the probability that pixel i is predicted as class t by the DAMF-Net network, represents the true probability that pixel i belongs to the tth class, N C Indicates the total number of categories to be segmented, N P represents the number of pixels in the image, α t and γ t is a weight parameter used to mitigate the impact of organ pixel ratio imbalance;

[0033] (4) Use the training dataset A to train the DAMF-Net network until the loss function L converges;

[0034] (5) Use the trained network to test the abdominal CT images and obtain the multi-organ segmentation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Schematic diagram of the DAMF-Net network structure according to an embodiment of the present invention

[0036] Figure 2 Schematic diagram of the coding block structure according to an embodiment of the present invention

[0037] Figure 3 Schematic diagram of the decoding block structure of an embodiment of the present invention

[0038] Figure 4 Schematic diagram of residual block structure according to an embodiment of the present invention

[0039] Figure 5 Schematic diagram of the multi-scale information fusion module structure of an embodiment of the present invention

[0040] Figure 6Schematic diagram of the anisotropic strip pooling module structure according to an embodiment of the present invention

[0041] Figure 7 Schematic diagram of the dual self-attention module structure of an embodiment of the present invention

[0042] Figure 8 Some experimental results obtained by the embodiments of the present invention are shown in the following examples: Figure 8 (a)~ Figure 8 (d) are four original CT images randomly selected from the FLARE dataset. Figure 8 (e)~ Figure 8 (h) is to use the embodiment of the present invention to Figure 8 (a)~ Figure 8 (d) Results of the tests DETAILED DESCRIPTION

[0043] Example 1

[0044] A multi-organ segmentation method for abdominal CT images based on deep learning, the specific implementation steps are as follows:

[0045] (1) Establish a training dataset A containing abdominal CT images and their corresponding multi-organ segmentation results;

[0046] (2) Construct a segmentation network based on dual self-attention mechanism and multi-scale feature fusion, called DAMF-Net, with the following structure: Figure 1 As shown, specifically including:

[0047] (2-a) A U-shaped structure with skip connections is used as the network benchmark framework, in which the encoding path and decoding path are composed of 5 cascaded encoding blocks and decoding blocks respectively. In order to avoid redundant information transmission and enhance the recognition ability of small targets and edges, a skip connection with a residual block is added between the encoding and decoding blocks of the same level. In addition, in order to compensate for the information loss caused by the limited receptive field, establish the connection between global information, and lightweight the network, a feature extraction module consisting of a multi-scale information fusion module and a dual self-attention module in parallel is introduced at the deepest layer of the network. The multi-scale information fusion module is abbreviated as MSCF and the dual self-attention module is abbreviated as DSA.

[0048] (2-b) The coding block described in step (2-a) has the following structure: Figure 2As shown, it is characterized by including a main encoding path and a residual branch. In the main encoding path, the input feature map is first subjected to two 3×3 convolutions. The first 3×3 convolution layer is sequentially connected to a normalization layer and a ReLU activation layer, and the second 3×3 convolution layer is connected to a normalization layer. To prevent network degradation, a residual branch is added to the encoding block to perform 1×1 convolution and normalization on the input features. Then, the results of the residual branch and the main encoding path are added. Finally, the output of the encoding block is obtained through a ReLU activation layer and a 2×2 maximum pooling layer.

[0049] (2-c) The decoding block described in step (2-a) has the following structure: Figure 3 As shown, it is characterized by including a decoding main path and a residual branch. In the decoding main path, the input feature map is first subjected to two 3×3 convolutions. The first 3×3 convolution layer is sequentially connected to a normalization layer and a ReLU activation layer, and the second 3×3 convolution layer is connected to a normalization layer. To prevent network degradation, a residual branch is added to the encoding block to perform 1×1 convolution and normalization on the input features. Then, the residual branch is added to the result of the decoding main path, and finally the output of the encoding block is obtained through a ReLU activation layer and upsampling, where the upsampling is achieved by 2×2 deconvolution.

[0050] (2-d) The residual block described in step (2-a) has the following structure: Figure 4 As shown, the specific structure includes: taking the output of the current level and the next level coding block as input, wherein the output of the current level coding block is recorded as F i , the output of the next level coding block is recorded as F i+1 ; The residual block first converts F i+1 Perform 2×2 deconvolution and ReLU activation operations to obtain the same value as F i Feature maps F of the same size up , and F i With F up Subtract and obtain the residual feature map F containing small targets and edge details sub =F i -F up ; In order to further enhance the expression of residual features, F sub Perform two consecutive 3×3 convolution operations, and connect a normalization layer and a ReLU activation layer in sequence after each 3×3 convolution layer;

[0051] (2-e) The multi-scale information fusion module described in step (2-a), namely, MSCF, has the following structure: Figure 5 As shown, the specific structure includes: first, the input feature F in Perform 1×1 convolution, normalization, and ReLU activation operations in sequence to obtain feature Fconv1×1 The purpose of this operation is to reduce the number of channels, reduce the number of parameters, and speed up the calculation process in the subsequent modules; then, F conv1×1 Input them into the pyramid pooling module to obtain the features F containing local information MS1 , and F conv1×1 Input to the anisotropic strip pooling module to obtain the feature F containing long-distance context information MS2 ; F MS1 and F MS2 Perform addition fusion, and then perform 1×1 convolution and normalization operations in sequence to obtain F that fuses multi-scale feature information MSCF ; In order to ensure the semantic consistency, the input feature F in With F MSCF The sum is added and output through the ReLU activation layer, and then 3×3 convolution, normalization, and ReLU activation operations are performed in sequence. Finally, the channel dimension is restored to the size of the input feature through 1×1 convolution to obtain the output feature F out ;

[0052] (2-f) The pyramid pooling module described in step (2-e) is denoted as PPM, which is characterized by the following structure: first, the input feature F conv1×1 Perform average pooling with kernels of 3×3, 5×5, and 7×7 respectively to obtain feature maps F of different scales pool3×3 、F pool5×5 and F pool7×7 , and respectively for F pool3×3 、F pool5×5 and F pool7×7 Perform 3×3 convolution and normalization in sequence to obtain the feature map F p1 、F p2 、F p3 , then, respectively for F p1 、F p2 、F p3 Perform deconvolution with kernels of 3×3, 5×5, and 7×7 to obtain features F1, F2, and F3; finally, add and fuse F1, F2, and F3, and perform ReLU activation, 3×3 convolution, normalization, and ReLU activation in sequence to obtain feature F MS1 :

[0053] (2-g) The anisotropic strip pooling module described in step (2-e) is denoted as ASPM, and its structure is as follows: Figure 6 As shown, the specific structure includes: first, using the pooling kernels of size H×1 and 1×W to filter F in the vertical and horizontal directions respectively. conv1×1 Perform average pooling to generate feature map F h and F v, where H and W represent the input feature map F conv1×1 The height and width of F h Perform 1×3 convolution and normalization in sequence to obtain the feature F′ h , for F v Perform 3×1 convolution and normalization in sequence to obtain the feature F′ v , and the obtained feature F′ h and F′ v Copy and expand along the vertical and horizontal directions respectively to obtain the same feature as the input F conv1×1 The feature maps F1′ and F2′ of the same size are then added and fused to obtain the feature F containing the long-range context information relationship. fusion ; Finally, for F fusion Perform ReLU activation, 3×3 convolution, normalization, and ReLU activation in sequence to obtain feature F MS2 ;

[0054] (2-h) The dual self-attention module described in step (2-a) is denoted as DSA, and its structure is as follows Figure 7 As shown, the specific structure includes: a position attention module and a channel attention module in parallel, which inputs the feature F down Input the position attention module and channel attention module respectively to obtain the feature F PAM and F CAM , and then add and fuse them to get the output feature

[0055] (2-i) The position attention module described in step (2-h), denoted as PAM, is characterized by a specific structure including:

[0056] (2-i-Ⅰ) For input features Where W′ and H′ are the width and height of the feature map, C is the number of channels of the feature map, and F is calculated on the three branches. down Perform shift convolution to obtain features and

[0057] (2-i-Ⅱ) In order to obtain the mutual dependence between pixels, the reshape operation is used to transform F p1 and F p2 Reconstructed into and And F′ p1 and F′ p2 Perform matrix multiplication and Softmax activation to obtain the weight matrix

[0058] (2-i-Ⅲ) will Refactored to And F′ p3 With M PAM Perform matrix multiplication to obtain

[0059] (2-i-Ⅳ)will Refactored to And F′ pmap With input feature F down Perform additive fusion and obtain output features through shift convolution

[0060] (2-j) The channel attention module described in step (2-h), denoted as CAM, is characterized by its overall structure being similar to that of PAM, specifically comprising:

[0061] (2-j-Ⅰ) Use reshape operation to reshape F on three branches respectively. down Reconstructed into

[0062] (2-j-Ⅱ) for F c1 Transpose and obtain And F′ c1 and F c2 Perform matrix multiplication and Softmax activation to obtain the weight matrix

[0063] (2-j-Ⅲ) F c3 With M CAM Perform matrix multiplication to obtain

[0064] (2-j-Ⅳ) Refactored to And F′ cmap With input feature F down Perform additive fusion and obtain output features through shift convolution

[0065] (3) Combining Dice loss and Focal loss to construct the loss function L of the DAMF-Net network:

[0066] L=L focal +λL dice

[0067]

[0068]

[0069]

[0070] Among them, L focal and Ldice Represent Focal loss and Dice loss respectively, λ is the weight coefficient, which controls the relative importance of Focal loss and Dice loss. In this embodiment, λ=0.4 is preferred. represents the probability that pixel i is predicted as class t by the DAMF-Net network, represents the true probability that pixel i belongs to the tth class, N C Indicates the total number of categories to be segmented, N P represents the number of pixels in the image, α t and γ t is a weight parameter used to reduce the impact of imbalanced organ pixel ratios. In this embodiment, α is preferred. t =1,γ t =2;

[0071] (4) Use the training dataset A to train the DAMF-Net network until the loss function L converges;

[0072] (5) Use the trained network to test the abdominal CT images and obtain the multi-organ segmentation results.

[0073] Example 2

[0074] The method described in Example 1 was used to conduct experiments on the FLARE public dataset. FLARE is a dataset that publishes 361 abdominal CT sequences and their corresponding gold-standard segmentation targets. The segmentation targets are primarily the liver, spleen, pancreas, and kidneys. In the experiment, these 361 CT sequences were divided into a test dataset and a training dataset in a ratio of 2:8.

[0075] The experimental environment of this embodiment is based on the Pytorch deep learning framework. All training is completed on an Nvidia GeForce RTX3060-12GB GPU. The initial learning rate is 0.005, the optimizer is set to Adam, and the learning rate is updated according to the ReduceLRPlateau strategy.

[0076] The Dice values of the liver, spleen, kidney, and pancreas obtained by the method of the present invention on the FLARE dataset are as high as 0.967, 0.944, 0.960, and 0.734, respectively. Figure 8 As shown, Figure 8 (a)~ Figure 8 (d) are four CT original images randomly selected from the test data. Figure 8 (e)~ Figure 8 (h) is to adopt the method in Example 1 to Figure 8 (a)~ Figure 8(d) The test results show that the method of the present invention can effectively segment organs such as the liver, spleen, kidney, and pancreas.

[0077] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-organ segmentation method for abdominal CT images based on deep learning, characterized in that: The following steps are involved: (1) Establish a training dataset A containing abdominal CT images and their corresponding multi-organ segmentation results; (2) Construct a segmentation network based on dual self-attention mechanism and multi-scale feature fusion, called DAMF-Net, which specifically includes: (2-a) A U-shaped structure with skip connections is used as the network benchmark framework, in which the encoding path and decoding path are composed of 5 cascaded encoding blocks and decoding blocks respectively. In order to avoid redundant information transmission and enhance the recognition ability of small targets and edges, a skip connection with a residual block is added between the encoding and decoding blocks of the same level. In addition, in order to compensate for the information loss caused by the limited receptive field, establish the connection between global information, and lightweight the network, a feature extraction module consisting of a multi-scale information fusion module and a dual self-attention module in parallel is introduced at the deepest layer of the network. The multi-scale information fusion module is abbreviated as MSCF and the dual self-attention module is abbreviated as DSA. (2-b) The coding block described in step (2-a) is characterized in that it includes a coding main path and a residual branch, and in the coding main path, the input feature map is first subjected to two 3×3 convolutions in sequence, the first 3×3 convolution layer is sequentially connected to a normalization layer and a ReLU activation layer, and the second 3×3 convolution layer is sequentially connected to a normalization layer; in order to prevent network degradation, a residual branch is added to the coding block to perform 1×1 convolution and normalization on the input features; then, the results of the residual branch and the coding main path are added, and finally, the output of the coding block is obtained through a ReLU activation layer and a 2×2 maximum pooling layer; (2-c) The decoding block described in step (2-a) is characterized in that it includes a decoding main path and a residual branch, and in the decoding main path, the input feature map is first subjected to two 3×3 convolutions in sequence, the first 3×3 convolution layer is connected in sequence to a normalization layer and a ReLU activation layer, and the second 3×3 convolution layer is connected to a normalization layer; in order to prevent network degradation, a residual branch is added to the encoding block, and the input features are subjected to 1×1 convolution and normalization; then, the residual branch is added to the result of the decoding main path, and finally the output of the encoding block is obtained through a ReLU activation layer and upsampling, wherein the upsampling is implemented by 2×2 deconvolution; (2-d) The residual block described in step (2-a) is characterized in that the specific structure includes: taking the output of the current level and the next level coding block as input, wherein the output of the current level coding block is denoted as F i , the output of the next level coding block is recorded as F i+1 ; The residual block first converts F i+1 Perform 2×2 deconvolution and ReLU activation operations to obtain the same value as F i Feature maps F of the same size up , and F i With F up Subtract and obtain the residual feature map F containing small targets and edge details sub =F i -F up ; In order to further enhance the expression of residual features, F sub Perform two consecutive 3×3 convolution operations, and connect a normalization layer and a ReLU activation layer in sequence after each 3×3 convolution layer; (2-e) The multi-scale information fusion module described in step (2-a), namely, MSCF, is characterized by the following specific structure: first, the input feature F in Perform 1×1 convolution, normalization, and ReLU activation operations in sequence to obtain feature F conv1×1 The purpose of this operation is to reduce the number of channels, reduce the number of parameters, and speed up the calculation process in the subsequent modules; then, F conv1×1 Input them into the pyramid pooling module to obtain the features F containing local information MS1 , and F conv1×1 Input to the anisotropic strip pooling module to obtain the feature F containing long-distance context information MS2 ; F MS1 and F MS2 Perform addition fusion, and then perform 1×1 convolution and normalization operations in sequence to obtain F that fuses multi-scale feature information MSCF ; In order to ensure the semantic consistency, the input feature F in With F MSCF The sum is added and output through the ReLU activation layer, and then 3×3 convolution, normalization, and ReLU activation operations are performed in sequence. Finally, the channel dimension is restored to the size of the input feature through 1×1 convolution to obtain the output feature F out ; (2-f) The pyramid pooling module described in step (2-e) is denoted as PPM, which is characterized by the following specific structure: first, the input feature F conv1×1 Perform average pooling with kernels of 3×3, 5×5, and 7×7 respectively to obtain feature maps F of different scales pool3×3 、F pool5×5 and F pool7×7 , and respectively for F pool3×3 、F pool5×5 and F pool7×7 Perform 3×3 convolution and normalization in sequence to obtain the feature map F p1 、F p2 、F p3 , then, respectively for F p1 、F p2 、F p3 Perform deconvolution with kernels of 3×3, 5×5, and 7×7 to obtain features F1, F2, and F3; finally, add and fuse F1, F2, and F3, and perform ReLU activation, 3×3 convolution, normalization, and ReLU activation in sequence to obtain feature F MS1 : (2-g) The anisotropic strip pooling module described in step (2-e) is denoted as ASPM, which is characterized by the following structure: first, using pooling kernels of size H×1 and 1×W to pool F in the vertical and horizontal directions respectively. conv1×1 Perform average pooling to generate feature map F h and F v , where H and W represent the input feature map F conv1×1 The height and width of F h Perform 1×3 convolution and normalization in sequence to obtain the feature F′ h , for F v Perform 3×1 convolution and normalization in sequence to obtain the feature F′ v , and the obtained feature F′ h and F′ v Copy and expand along the vertical and horizontal directions respectively to obtain the same feature as the input F conv1×1 The feature maps F1′ and F2′ of the same size are then added and fused to obtain the feature F containing the long-range context information relationship. fusion ; Finally, for F fusion Perform ReLU activation, 3×3 convolution, normalization, and ReLU activation in sequence to obtain feature F MS2 ; (2-h) The dual self-attention module described in step (2-a) is denoted as DSA, which is characterized by a specific structure comprising: a position attention module and a channel attention module in parallel, which takes the input feature F down Input the position attention module and channel attention module respectively to obtain the feature F PAM and F CAM , and then add and fuse them to get the output feature (2-i) The position attention module described in step (2-h), denoted as PAM, is characterized by a specific structure including: (2-i-Ⅰ) For input features Where W′ and H′ are the width and height of the feature map, C is the number of channels of the feature map, and F is calculated on the three branches. down Perform shift convolution to obtain features and (2-i-Ⅱ) In order to obtain the mutual dependence between pixels, the reshape operation is used to transform F p1 and F p2 Reconstructed into and And F′ p1 and F′ p2 Perform matrix multiplication and Softmax activation to obtain the weight matrix (2-i-Ⅲ) will Refactored to And F p ′3 and M PAM Perform matrix multiplication to obtain (2-i-Ⅳ)will Refactored to And F p ' map With input feature F down Perform additive fusion and obtain output features through shift convolution (2-j) The channel attention module described in step (2-h), denoted as CAM, is characterized by its overall structure being similar to that of PAM, specifically comprising: (2-j-Ⅰ) Use reshape operation to reshape F on three branches respectively. down Reconstructed into (2-j-Ⅱ) for F c1 Transpose and obtain And F c ′1 and F c2 Perform matrix multiplication and Softmax activation to obtain the weight matrix (2-j-Ⅲ) F c3 With M CAM Perform matrix multiplication to obtain (2-j-Ⅳ) Refactored to And F c ' map With input feature F down Perform additive fusion and obtain output features through shift convolution (3) Combining Dice loss and Focal loss to construct the loss function L of the DAMF-Net network: L=L focal +λL dice Among them, L focal and L dice Represent Focal loss and Dice loss respectively, λ is the weight coefficient, which controls the relative importance of Focal loss and Dice loss. represents the probability that pixel i is predicted as class t by the DAMF-Net network, represents the true probability that pixel i belongs to the tth class, N C Indicates the total number of categories to be segmented, N P represents the number of pixels in the image, α t and γ t is a weight parameter used to mitigate the impact of organ pixel ratio imbalance; (4) Use the training dataset A to train the DAMF-Net network until the loss function L converges; (5) Use the trained network to test the abdominal CT images and obtain the multi-organ segmentation results.

2. The deep learning-based multi-organ segmentation method for abdominal CT images according to claim 1, characterized in that: The loss function weight coefficient λ described in step (3) is preferably a constant between 0.3 and 0.7.

Images